Mercurial > repos > goeckslab > bagging_tool
comparison mil_bag.py @ 0:e6e9ea0703ef draft default tip
planemo upload for repository https://github.com/goeckslab/gleam.git commit 783551569c645073698fce50f1ed9c4605b3e65a
| author | goeckslab |
|---|---|
| date | Thu, 19 Jun 2025 23:31:55 +0000 |
| parents | |
| children |
comparison
equal
deleted
inserted
replaced
| -1:000000000000 | 0:e6e9ea0703ef |
|---|---|
| 1 """ | |
| 2 A script for creating bags of instances from embeddings | |
| 3 and metadata for Multiple Instance Learning (MIL) tasks. | |
| 4 | |
| 5 Processes embedding and metadata CSV files to generate | |
| 6 bags of instances, saved as a single CSV file. Supports | |
| 7 bagging strategies (by sample, in turns, or random), | |
| 8 pooling methods, and options for balancing, preventing | |
| 9 data leakage, and Ludwig formatting. Handles large | |
| 10 datasets efficiently using temporary Parquet files, | |
| 11 sequential processing, and multiprocessing. | |
| 12 | |
| 13 Dependencies: | |
| 14 - gc: For manual garbage collection to manage memory. | |
| 15 - argparse: For parsing command-line arguments. | |
| 16 - logging: For logging progress and errors. | |
| 17 - multiprocessing (mp): For parallel processing. | |
| 18 - os: For file operations and temporary file management. | |
| 19 - tempfile: For creating temporary files. | |
| 20 - numpy (np): For numerical operations and array. | |
| 21 - pandas (pd): For data manipulation and I/O (CSV, Parquet). | |
| 22 - torch: For tensor operations (attention pooling). | |
| 23 - torch.nn: For NN components (attention pooling). | |
| 24 - fastparquet: For reading and writing Parquet files. | |
| 25 | |
| 26 Key Features: | |
| 27 - Multiple bagging: by sample (`bag_by_sample`), in | |
| 28 turns (`bag_in_turns`), or random (`bag_random`). | |
| 29 - Various pooling methods (e.g., max, mean, attention). | |
| 30 - Prevents data leakage by splitting at sample level. | |
| 31 - Balances bags by label imbalance or truncating. | |
| 32 - Outputs in Ludwig format (whitespace-separated vectors). | |
| 33 - Efficient large dataset processing (temp Parquet, | |
| 34 sequential CSV write). | |
| 35 - GPU acceleration for certain pooling (e.g., attention). | |
| 36 | |
| 37 Usage: | |
| 38 Run the script from the command line with arguments: | |
| 39 | |
| 40 ```bash | |
| 41 python ludwig_mil_temp.py --embeddings_csv <path_to_embeddings.csv> | |
| 42 --metadata_csv <path_to_metadata.csv> --bag_size <bag_size> | |
| 43 --pooling_method <method> --output_csv <output.csv> | |
| 44 [--split_proportions <train,val,test>] [--dataleak] | |
| 45 [--balance_enforced] [--by_sample <splits>] [--repeats <num>] | |
| 46 [--ludwig_format] [--random_seed <seed>] | |
| 47 [--imbalance_cap <percentage>] [--truncate_bags] [--use_gpu] | |
| 48 """ | |
| 49 | |
| 50 import argparse | |
| 51 import gc | |
| 52 import logging | |
| 53 import multiprocessing as mp | |
| 54 import os | |
| 55 import tempfile | |
| 56 | |
| 57 import numpy as np | |
| 58 import pandas as pd | |
| 59 import torch | |
| 60 import torch.nn as nn | |
| 61 | |
| 62 | |
| 63 def parse_bag_size(bag_size_str): | |
| 64 """Parses bag size string into a range or single value.""" | |
| 65 try: | |
| 66 if '-' in bag_size_str: | |
| 67 start, end = map(int, bag_size_str.split('-')) | |
| 68 return list(range(start, end + 1)) | |
| 69 return [int(bag_size_str)] | |
| 70 except ValueError: | |
| 71 logging.error("Invalid bag_size format: %s", bag_size_str) | |
| 72 raise | |
| 73 | |
| 74 | |
| 75 def parse_by_sample(value): | |
| 76 """Parses by_sample string into a set of split values.""" | |
| 77 try: | |
| 78 value = str(value) | |
| 79 splits = [int(x) for x in value.split(",")] | |
| 80 valid_splits = {0, 1, 2} | |
| 81 if not all(x in valid_splits for x in splits): | |
| 82 logging.warning("Invalid splits in by_sample: %s", splits) | |
| 83 return None | |
| 84 return splits | |
| 85 except (ValueError, AttributeError): | |
| 86 logging.warning("By_Sample not used") | |
| 87 return None | |
| 88 | |
| 89 | |
| 90 class BaggingConfig: | |
| 91 """Configuration class for bagging parameters.""" | |
| 92 | |
| 93 def __init__(self, params): | |
| 94 self.embeddings_csv = params.embeddings_csv | |
| 95 self.metadata_csv = params.metadata_csv | |
| 96 self.split_proportions = params.split_proportions | |
| 97 self.prevent_leakage = params.dataleak | |
| 98 self.balance_enforced = params.balance_enforced | |
| 99 self.bag_size = parse_bag_size(params.bag_size) | |
| 100 self.pooling_method = params.pooling_method | |
| 101 self.by_sample = parse_by_sample(params.by_sample) | |
| 102 self.repeats = params.repeats | |
| 103 self.ludwig_format = params.ludwig_format | |
| 104 self.output_csv = params.output_csv | |
| 105 self.random_seed = params.random_seed | |
| 106 self.imbalance_cap = params.imbalance_cap | |
| 107 self.truncate_bags = params.truncate_bags | |
| 108 self.use_gpu = params.use_gpu | |
| 109 | |
| 110 def __str__(self): | |
| 111 """String representation of the config for logging.""" | |
| 112 return ( | |
| 113 f"embeddings_csv={self.embeddings_csv}, " | |
| 114 f"metadata_csv={self.metadata_csv}, " | |
| 115 f"split_proportions={self.split_proportions}, " | |
| 116 f"prevent_leakage={self.prevent_leakage}, " | |
| 117 f"balance_enforced={self.balance_enforced}, " | |
| 118 f"bag_size={self.bag_size}, " | |
| 119 f"pooling_method={self.pooling_method}, " | |
| 120 f"by_sample={self.by_sample}, " | |
| 121 f"repeats={self.repeats}, " | |
| 122 f"ludwig_format={self.ludwig_format}, " | |
| 123 f"output_csv={self.output_csv}, " | |
| 124 f"random_seed={self.random_seed}, " | |
| 125 f"imbalance_cap={self.imbalance_cap}, " | |
| 126 f"truncate_bags={self.truncate_bags}, " | |
| 127 f"use_gpu={self.use_gpu}" | |
| 128 ) | |
| 129 | |
| 130 | |
| 131 def set_random_seed(configs): | |
| 132 """Sets random seeds for reproducibility.""" | |
| 133 np.random.seed(configs.random_seed) | |
| 134 torch.manual_seed(configs.random_seed) | |
| 135 if torch.cuda.is_available(): | |
| 136 torch.cuda.manual_seed_all(configs.random_seed) | |
| 137 torch.backends.cudnn.deterministic = True | |
| 138 torch.backends.cudnn.benchmark = False | |
| 139 logging.info("Random seed set to %d", configs.random_seed) | |
| 140 | |
| 141 | |
| 142 def validate_metadata(metadata): | |
| 143 """Validates metadata for required columns.""" | |
| 144 required_cols = {"sample_name", "label"} | |
| 145 if not required_cols.issubset(metadata.columns): | |
| 146 missing = required_cols - set(metadata.columns) | |
| 147 raise ValueError(f"Metadata missing columns: {missing}") | |
| 148 return metadata | |
| 149 | |
| 150 | |
| 151 def load_metadata(file_path): | |
| 152 """Loads metadata from a CSV file.""" | |
| 153 metadata = pd.read_csv(file_path) | |
| 154 validate_metadata(metadata) | |
| 155 logging.info("Metadata loaded with %d samples, cols: %s", | |
| 156 len(metadata), list(metadata.columns)) | |
| 157 logging.info("Unique samples: %d, labels: %d", | |
| 158 metadata["sample_name"].nunique(), | |
| 159 metadata["label"].nunique()) | |
| 160 return metadata | |
| 161 | |
| 162 | |
| 163 def convert_proportions(proportion_string): | |
| 164 """Converts a string of split proportions into a list of floats.""" | |
| 165 proportion_list = [float(p) for p in proportion_string.split(",")] | |
| 166 print(proportion_list) | |
| 167 if len(proportion_list) == 2: | |
| 168 proportion_list = [proportion_list[0], 0.0, proportion_list[1]] | |
| 169 | |
| 170 for proportion in proportion_list: | |
| 171 if proportion < 0 or proportion > 1: | |
| 172 raise ValueError("Each proportion must be between 0 and 1") | |
| 173 | |
| 174 if abs(sum(proportion_list) - 1.0) > 1e-6: | |
| 175 raise ValueError("Proportions must sum to approximately 1.0") | |
| 176 | |
| 177 return proportion_list | |
| 178 | |
| 179 | |
| 180 def calculate_split_counts(total_samples, proportions): | |
| 181 """Calculates sample counts for each split.""" | |
| 182 counts = [int(p * total_samples) for p in proportions] | |
| 183 calculated_total = sum(counts) | |
| 184 if calculated_total < total_samples: | |
| 185 counts[-1] += total_samples - calculated_total | |
| 186 elif calculated_total > total_samples: | |
| 187 counts[0] -= calculated_total - total_samples | |
| 188 return counts | |
| 189 | |
| 190 | |
| 191 def assign_split_labels(proportions, sample_count): | |
| 192 """Assigns split labels based on proportions.""" | |
| 193 proportion_values = convert_proportions(proportions) | |
| 194 train_fraction, val_fraction, test_fraction = proportion_values | |
| 195 | |
| 196 if val_fraction == 0 and test_fraction == 0: | |
| 197 labels = np.zeros(sample_count, dtype=int) | |
| 198 elif val_fraction == 0: | |
| 199 train_size = int(train_fraction * sample_count) | |
| 200 test_size = sample_count - train_size | |
| 201 labels = np.array([0] * train_size + [2] * test_size) | |
| 202 else: | |
| 203 split_counts = calculate_split_counts(sample_count, proportion_values) | |
| 204 labels = np.concatenate([ | |
| 205 np.zeros(split_counts[0], dtype=int), | |
| 206 np.ones(split_counts[1], dtype=int), | |
| 207 2 * np.ones(split_counts[2], dtype=int) | |
| 208 ]) | |
| 209 return labels | |
| 210 | |
| 211 | |
| 212 def split_dataset(metadata, configs): | |
| 213 """Splits dataset into train, val, test sets if prevent_leakage is True.""" | |
| 214 if configs.prevent_leakage: | |
| 215 logging.info("No data leakage allowed") | |
| 216 unique_samples = metadata["sample_name"].unique() | |
| 217 sample_count = len(unique_samples) | |
| 218 split_labels = assign_split_labels(configs.split_proportions, | |
| 219 sample_count) | |
| 220 shuffled_samples = np.random.permutation(unique_samples) | |
| 221 label_series = pd.Series(split_labels, index=shuffled_samples) | |
| 222 metadata["split"] = metadata["sample_name"].map(label_series) | |
| 223 train_count = (metadata["split"] == 0).sum() | |
| 224 val_count = (metadata["split"] == 1).sum() | |
| 225 test_count = (metadata["split"] == 2).sum() | |
| 226 logging.info("Dataset split: train %d, val %d, test %d", | |
| 227 train_count, val_count, test_count) | |
| 228 else: | |
| 229 logging.info("Data leakage allowed setup") | |
| 230 return metadata | |
| 231 | |
| 232 | |
| 233 def assign_chunk_splits(chunk, split_counts, current_counts): | |
| 234 """Assigns split labels to a chunk of embeddings.""" | |
| 235 chunk_size = len(chunk) | |
| 236 remaining = { | |
| 237 0: split_counts[0] - current_counts[0], | |
| 238 1: split_counts[1] - current_counts[1], | |
| 239 2: split_counts[2] - current_counts[2] | |
| 240 } | |
| 241 available_splits = [s for s, count in remaining.items() if count > 0] | |
| 242 if not available_splits: | |
| 243 return chunk, current_counts | |
| 244 | |
| 245 total_remaining = sum(remaining.values()) | |
| 246 assign_count = min(chunk_size, total_remaining) | |
| 247 if assign_count == 0: | |
| 248 return chunk, current_counts | |
| 249 | |
| 250 weights = [remaining[s] / total_remaining for s in available_splits] | |
| 251 splits = np.random.choice(available_splits, size=assign_count, p=weights) | |
| 252 chunk["split"] = pd.Series(splits, index=chunk.index[:assign_count]) | |
| 253 chunk["split"] = chunk["split"].fillna(0).astype(int) | |
| 254 | |
| 255 for split in available_splits: | |
| 256 current_counts[split] += np.sum(splits == split) | |
| 257 | |
| 258 return chunk, current_counts | |
| 259 | |
| 260 | |
| 261 def setup_temp_files(): | |
| 262 """Sets up temporary Parquet files for splits and bag outputs.""" | |
| 263 splits = [0, 1, 2] | |
| 264 split_files = {} | |
| 265 for split in splits: | |
| 266 fd, path = tempfile.mkstemp(prefix=f"split_{split}_", | |
| 267 suffix=".parquet", | |
| 268 dir=os.getcwd()) | |
| 269 os.close(fd) # Explicitly close the file descriptor | |
| 270 split_files[split] = path | |
| 271 | |
| 272 bag_outputs = {} | |
| 273 for split in splits: | |
| 274 fd, path = tempfile.mkstemp(prefix=f"MIL_bags_{split}_", | |
| 275 suffix=".parquet", | |
| 276 dir=os.getcwd()) | |
| 277 os.close(fd) # Explicitly close the file descriptor | |
| 278 bag_outputs[split] = path | |
| 279 | |
| 280 return split_files, bag_outputs | |
| 281 | |
| 282 | |
| 283 def distribute_embeddings(configs, metadata, split_files): | |
| 284 embeddings_path = configs.embeddings_csv | |
| 285 proportion_string = configs.split_proportions | |
| 286 prevent_leakage = configs.prevent_leakage | |
| 287 | |
| 288 logging.info("Distributing embeddings from %s to Parquet files", | |
| 289 embeddings_path) | |
| 290 buffer_size = 50000 | |
| 291 merged_header = None | |
| 292 non_sample_columns = None | |
| 293 | |
| 294 if not prevent_leakage: | |
| 295 logging.warning( | |
| 296 "Counting rows in %s; may be slow for large files", | |
| 297 embeddings_path | |
| 298 ) | |
| 299 total_rows = sum(1 for _ in open(embeddings_path)) - 1 | |
| 300 proportions = convert_proportions(proportion_string) | |
| 301 split_counts = calculate_split_counts(total_rows, proportions) | |
| 302 current_counts = {0: 0, 1: 0, 2: 0} | |
| 303 else: | |
| 304 sample_to_split = dict(zip(metadata["sample_name"], metadata["split"])) | |
| 305 sample_to_label = dict(zip(metadata["sample_name"], metadata["label"])) | |
| 306 | |
| 307 first_write = {split: True for split in split_files} | |
| 308 | |
| 309 try: | |
| 310 first_header_read = True | |
| 311 for chunk in pd.read_csv(embeddings_path, chunksize=buffer_size): | |
| 312 # Modify 'sample_name' to remove part after the last underscore | |
| 313 chunk['sample_name'] = chunk['sample_name'].apply(lambda x: x.rsplit('_', 1)[0]) | |
| 314 | |
| 315 if first_header_read: | |
| 316 orig_header = list(chunk.columns) | |
| 317 non_sample_columns = [ | |
| 318 col for col in orig_header if col != "sample_name" | |
| 319 ] | |
| 320 merged_header = ["sample_name", "label"] + non_sample_columns | |
| 321 logging.info("Merged header: %s", merged_header) | |
| 322 first_header_read = False | |
| 323 | |
| 324 if prevent_leakage: | |
| 325 chunk["split"] = chunk["sample_name"].map(sample_to_split) | |
| 326 chunk["label"] = chunk["sample_name"].map(sample_to_label) | |
| 327 else: | |
| 328 chunk, current_counts = assign_chunk_splits(chunk, | |
| 329 split_counts, | |
| 330 current_counts) | |
| 331 chunk = chunk.merge(metadata[["sample_name", "label"]], | |
| 332 on="sample_name", | |
| 333 how="left") | |
| 334 | |
| 335 chunk = chunk.dropna(subset=["split", "label"]) | |
| 336 for split in split_files: | |
| 337 split_chunk = chunk[chunk["split"] == split] | |
| 338 if not split_chunk.empty: | |
| 339 temp_file = split_files[split] | |
| 340 split_chunk[merged_header].to_parquet( | |
| 341 temp_file, | |
| 342 engine="fastparquet", | |
| 343 append=not first_write[split], | |
| 344 index=False | |
| 345 ) | |
| 346 first_write[split] = False | |
| 347 del chunk | |
| 348 gc.collect() | |
| 349 | |
| 350 except Exception as e: | |
| 351 logging.error("Error distributing embeddings to Parquet: %s", e) | |
| 352 raise | |
| 353 | |
| 354 | |
| 355 def aggregate_embeddings(embeddings, pooling_method, use_gpu=False): | |
| 356 # Convert embeddings to a float32 array explicitly. | |
| 357 embeddings = np.asarray(embeddings, dtype=np.float32) | |
| 358 | |
| 359 if embeddings.ndim == 1: | |
| 360 embeddings = embeddings.reshape(1, -1) | |
| 361 elif embeddings.ndim == 0: | |
| 362 embeddings = embeddings.reshape(1, 1) | |
| 363 | |
| 364 logging.debug("Aggregating embeddings with shape: %s", embeddings.shape) | |
| 365 | |
| 366 if pooling_method == "max_pooling": | |
| 367 result = np.max(embeddings, axis=0) | |
| 368 elif pooling_method == "mean_pooling": | |
| 369 result = np.mean(embeddings, axis=0) | |
| 370 elif pooling_method == "sum_pooling": | |
| 371 result = np.sum(embeddings, axis=0) | |
| 372 elif pooling_method == "min_pooling": | |
| 373 result = np.min(embeddings, axis=0) | |
| 374 elif pooling_method == "median_pooling": | |
| 375 result = np.median(embeddings, axis=0) | |
| 376 elif pooling_method == "l2_norm_pooling": | |
| 377 norm = np.linalg.norm(embeddings, axis=1, keepdims=True) | |
| 378 if norm.any(): | |
| 379 result = np.mean(embeddings / (norm + 1e-8), axis=0) | |
| 380 else: | |
| 381 result = np.mean(embeddings, axis=0) | |
| 382 elif pooling_method == "geometric_mean_pooling": | |
| 383 clipped = np.clip(embeddings, 1e-10, None) | |
| 384 result = np.exp(np.mean(np.log(clipped), axis=0)) | |
| 385 elif pooling_method == "first_embedding": | |
| 386 result = embeddings[0] | |
| 387 elif pooling_method == "last_embedding": | |
| 388 result = embeddings[-1] | |
| 389 elif pooling_method == "attention_pooling": | |
| 390 device = 'cuda' if use_gpu and torch.cuda.is_available() else 'cpu' | |
| 391 tensor = torch.tensor(embeddings, dtype=torch.float32).to(device) | |
| 392 with torch.no_grad(): | |
| 393 linear = nn.Linear(tensor.shape[1], 1).to(device) | |
| 394 weights = nn.Softmax(dim=0)(linear(tensor)) | |
| 395 result = torch.sum(weights * tensor, dim=0).cpu().detach().numpy() | |
| 396 else: | |
| 397 raise ValueError(f"Unknown pooling method: {pooling_method}") | |
| 398 | |
| 399 logging.debug("Aggregated embedding shape: %s", result.shape) | |
| 400 return result | |
| 401 | |
| 402 | |
| 403 def bag_by_sample(df, split, bag_file, config, batch_size=1000, | |
| 404 fixed_target_bags=None): | |
| 405 """ | |
| 406 Processes the provided DataFrame by grouping rows by sample, | |
| 407 constructs bags from each sample group using the configured bag_size, | |
| 408 and writes the bag rows directly to bag_file (a Parquet file) in batches. | |
| 409 | |
| 410 Args: | |
| 411 df (pd.DataFrame): The DataFrame containing the data. | |
| 412 split (str): The split identifier (e.g., 'train', 'val'). | |
| 413 bag_file (str): The path to the Parquet file to write the bags. | |
| 414 config (object): Configuration object with bag_size, pooling_method... | |
| 415 batch_size (int, optional): The number of rows to write in each batch. | |
| 416 fixed_target_bags (tuple, optional): (target_label, num_bags) | |
| 417 to generate bags only for target_label. | |
| 418 | |
| 419 Output row format: | |
| 420 sample_name, bag_label, split, bag_size, vector_0, vector_1, vector_N | |
| 421 """ | |
| 422 log_msg = f"Processing by sample for split: {split}" | |
| 423 if fixed_target_bags: | |
| 424 log_msg += f" with fixed target {fixed_target_bags}" | |
| 425 logging.info(log_msg) | |
| 426 | |
| 427 batch_rows = [] | |
| 428 bag_count = 0 | |
| 429 vector_columns = [ | |
| 430 col for col in df.columns | |
| 431 if col not in ["sample_name", "label", "split"] | |
| 432 ] | |
| 433 | |
| 434 if fixed_target_bags is not None: | |
| 435 target_label, target_needed = fixed_target_bags | |
| 436 target_samples = list( | |
| 437 df[df["label"] == target_label]["sample_name"].unique() | |
| 438 ) | |
| 439 df = df[df["sample_name"].isin(target_samples)] | |
| 440 | |
| 441 if df.empty: | |
| 442 logging.warning( | |
| 443 "No samples available for target label %d in split %s", | |
| 444 target_label, | |
| 445 split | |
| 446 ) | |
| 447 return | |
| 448 | |
| 449 available_samples = target_samples.copy() | |
| 450 np.random.shuffle(available_samples) | |
| 451 | |
| 452 while bag_count < target_needed: | |
| 453 if len(available_samples) == 0: | |
| 454 available_samples = target_samples.copy() | |
| 455 np.random.shuffle(available_samples) | |
| 456 logging.info( | |
| 457 "Reusing samples for target label %d in split %s", | |
| 458 target_label, | |
| 459 split | |
| 460 ) | |
| 461 | |
| 462 sample_name = available_samples.pop() | |
| 463 group = df[df["sample_name"] == sample_name] | |
| 464 embeddings = group[vector_columns].values | |
| 465 num_instances = len(group) | |
| 466 | |
| 467 current_bag_size = config.bag_size[0] \ | |
| 468 if len(config.bag_size) == 1 else \ | |
| 469 np.random.randint(config.bag_size[0], config.bag_size[1] + 1) | |
| 470 current_bag_size = min(current_bag_size, num_instances) | |
| 471 | |
| 472 selected = group.sample(n=current_bag_size, replace=True) | |
| 473 bag_embeddings = selected[vector_columns].values | |
| 474 | |
| 475 aggregated_embedding = aggregate_embeddings( | |
| 476 bag_embeddings, | |
| 477 config.pooling_method, | |
| 478 config.use_gpu | |
| 479 ) | |
| 480 | |
| 481 bag_label = int(any(selected["label"] == 1)) | |
| 482 if bag_label != target_label: | |
| 483 logging.warning( | |
| 484 "Generated bag for target %d but got label %d", | |
| 485 target_label, bag_label | |
| 486 ) | |
| 487 continue | |
| 488 | |
| 489 row = { | |
| 490 "sample_name": sample_name, | |
| 491 "bag_label": bag_label, | |
| 492 "split": split, | |
| 493 "bag_size": current_bag_size | |
| 494 } | |
| 495 for j, val in enumerate(aggregated_embedding): | |
| 496 row[f"vector_{j}"] = val | |
| 497 | |
| 498 batch_rows.append(row) | |
| 499 bag_count += 1 | |
| 500 | |
| 501 if len(batch_rows) >= batch_size: | |
| 502 df_batch = pd.DataFrame(batch_rows) | |
| 503 # Check if the file has data to determine append mode | |
| 504 append_mode = os.path.getsize(bag_file) > 0 | |
| 505 df_batch.to_parquet( | |
| 506 bag_file, | |
| 507 engine="fastparquet", | |
| 508 append=append_mode, | |
| 509 index=False | |
| 510 ) | |
| 511 logging.debug( | |
| 512 "Fixed mode: Wrote batch of %d rows to %s", | |
| 513 len(batch_rows), | |
| 514 bag_file | |
| 515 ) | |
| 516 batch_rows = [] | |
| 517 del df_batch | |
| 518 gc.collect() | |
| 519 | |
| 520 else: | |
| 521 # Standard mode: process all samples | |
| 522 groups = df.groupby("sample_name") | |
| 523 for sample_name, group in groups: | |
| 524 embeddings = group[vector_columns].values | |
| 525 labels = group["label"].values | |
| 526 num_instances = len(group) | |
| 527 | |
| 528 current_bag_size = config.bag_size[0] \ | |
| 529 if len(config.bag_size) == 1 else \ | |
| 530 np.random.randint( | |
| 531 config.bag_size[0], | |
| 532 config.bag_size[1] + 1 | |
| 533 ) | |
| 534 num_bags = ( | |
| 535 num_instances + current_bag_size - 1 | |
| 536 ) // current_bag_size | |
| 537 logging.info( | |
| 538 "Sample %s: %d instances, creating %d bags (bag size %d)", | |
| 539 sample_name, | |
| 540 num_instances, | |
| 541 num_bags, | |
| 542 current_bag_size | |
| 543 ) | |
| 544 | |
| 545 for i in range(num_bags): | |
| 546 start_idx = i * current_bag_size | |
| 547 end_idx = min(start_idx + current_bag_size, num_instances) | |
| 548 bag_embeddings = embeddings[start_idx:end_idx] | |
| 549 bag_labels = labels[start_idx:end_idx] | |
| 550 | |
| 551 aggregated_embedding = aggregate_embeddings( | |
| 552 bag_embeddings, | |
| 553 config.pooling_method, | |
| 554 config.use_gpu | |
| 555 ) | |
| 556 bag_label = int(any(bag_labels == 1)) | |
| 557 | |
| 558 row = { | |
| 559 "sample_name": sample_name, | |
| 560 "bag_label": bag_label, | |
| 561 "split": split, | |
| 562 "bag_size": end_idx - start_idx | |
| 563 } | |
| 564 for j, val in enumerate(aggregated_embedding): | |
| 565 row[f"vector_{j}"] = val | |
| 566 | |
| 567 batch_rows.append(row) | |
| 568 bag_count += 1 | |
| 569 | |
| 570 if len(batch_rows) >= batch_size: | |
| 571 df_batch = pd.DataFrame(batch_rows) | |
| 572 # Check if the file has data to determine append mode | |
| 573 append_mode = os.path.getsize(bag_file) > 0 | |
| 574 df_batch.to_parquet( | |
| 575 bag_file, | |
| 576 engine="fastparquet", | |
| 577 append=append_mode, | |
| 578 index=False | |
| 579 ) | |
| 580 logging.debug( | |
| 581 "Wrote batch of %d rows to %s", | |
| 582 len(batch_rows), | |
| 583 bag_file | |
| 584 ) | |
| 585 batch_rows = [] | |
| 586 del df_batch | |
| 587 gc.collect() | |
| 588 | |
| 589 # Write any remaining rows | |
| 590 if batch_rows: | |
| 591 df_batch = pd.DataFrame(batch_rows) | |
| 592 append_mode = os.path.getsize(bag_file) > 0 | |
| 593 df_batch.to_parquet( | |
| 594 bag_file, | |
| 595 engine="fastparquet", | |
| 596 append=append_mode, | |
| 597 index=False | |
| 598 ) | |
| 599 logging.debug( | |
| 600 "Wrote final batch of %d rows to %s", | |
| 601 len(batch_rows), | |
| 602 bag_file | |
| 603 ) | |
| 604 del df_batch | |
| 605 gc.collect() | |
| 606 | |
| 607 logging.info("Created %d bags for split: %s", bag_count, split) | |
| 608 | |
| 609 | |
| 610 def bag_in_turns(df, split, bag_file, config, batch_size=500, | |
| 611 fixed_target_bags=None, allow_reuse=True): | |
| 612 """ | |
| 613 Generate bags of instances from a DataFrame, with optional | |
| 614 fixed-target mode, data reuse, and enhanced diversity. | |
| 615 | |
| 616 Parameters: | |
| 617 - df (pd.DataFrame): Input DataFrame with columns including | |
| 618 'sample_name', 'label', 'split', and embedding vectors. | |
| 619 - split (str): Dataset split (e.g., 'train', 'test'). | |
| 620 - bag_file (str): Path to save the output Parquet file. | |
| 621 - config (object): Configuration object with attributes | |
| 622 'bag_size', 'pooling_method', and 'use_gpu'. | |
| 623 - batch_size (int): Number of bags to process before writing | |
| 624 to file (default: 500). | |
| 625 - fixed_target_bags (tuple): Optional (label, num_bags) to | |
| 626 generate bags for a specific label (e.g., (0, 100)). | |
| 627 - allow_reuse (bool): Allow resampling instances with | |
| 628 replacement if True (default: True). | |
| 629 | |
| 630 Returns: | |
| 631 - None: Saves bags to the specified Parquet file. | |
| 632 """ | |
| 633 logging.info( | |
| 634 "Processing bag in turns for split %s%s", | |
| 635 split, | |
| 636 (" with fixed target " + str(fixed_target_bags)) | |
| 637 if fixed_target_bags is not None else "" | |
| 638 ) | |
| 639 | |
| 640 # Identify embedding columns (exclude non-vector columns). | |
| 641 vector_columns = [ | |
| 642 col for col in df.columns | |
| 643 if col not in ["sample_name", "label", "split"] | |
| 644 ] | |
| 645 | |
| 646 # Convert the DataFrame to a NumPy array for faster processing. | |
| 647 df_np = df.to_numpy() | |
| 648 | |
| 649 # Determine bag size range from config. | |
| 650 if len(config.bag_size) == 1: | |
| 651 bag_min = bag_max = config.bag_size[0] | |
| 652 else: | |
| 653 bag_min, bag_max = config.bag_size | |
| 654 | |
| 655 batch_rows = [] | |
| 656 bag_count = 0 | |
| 657 | |
| 658 if fixed_target_bags is not None: | |
| 659 # Fixed-target mode: generate bags for a specific label. | |
| 660 target, target_needed = fixed_target_bags # e.g., (0, 100) | |
| 661 if target == 0: | |
| 662 # Optimize for target label 0: remove all label 1 instances | |
| 663 indices = np.where(df_np[:, 1] == 0)[0] | |
| 664 logging.info( | |
| 665 "Fixed mode: target label 0, using only label 0 instances, \ | |
| 666 total available %d rows", | |
| 667 len(indices) | |
| 668 ) | |
| 669 else: | |
| 670 # For target label 1, use all instances to allow mixing | |
| 671 indices = np.arange(len(df_np)) | |
| 672 logging.info( | |
| 673 "Fixed mode: target label 1, using all instances, \ | |
| 674 total available %d rows", | |
| 675 len(indices) | |
| 676 ) | |
| 677 | |
| 678 total_available = len(indices) | |
| 679 | |
| 680 while bag_count < target_needed: | |
| 681 current_bag_size = np.random.randint(bag_min, bag_max + 1) \ | |
| 682 if bag_min != bag_max else bag_min | |
| 683 | |
| 684 if total_available < current_bag_size and not allow_reuse: | |
| 685 logging.warning( | |
| 686 "Not enough instances (%d) for bag size %d and \ | |
| 687 target label %d", | |
| 688 total_available, current_bag_size, target | |
| 689 ) | |
| 690 break | |
| 691 | |
| 692 # Sample instances | |
| 693 selected = np.random.choice( | |
| 694 indices, | |
| 695 size=current_bag_size, | |
| 696 replace=allow_reuse | |
| 697 ) | |
| 698 bag_data = df_np[selected] | |
| 699 | |
| 700 if target == 1: | |
| 701 # For positive bags, ensure at least one instance has label 1 | |
| 702 if not np.any(bag_data[:, 1] == 1): | |
| 703 continue # Skip if no positive instance | |
| 704 bag_label = 1 | |
| 705 else: | |
| 706 # For negative bags, all instances are label 0 due to filtering | |
| 707 bag_label = 0 | |
| 708 | |
| 709 # Aggregate embeddings. | |
| 710 vec_col_indices = [ | |
| 711 df.columns.get_loc(col) for col in vector_columns | |
| 712 ] | |
| 713 embeddings = bag_data[:, vec_col_indices].astype(np.float32) | |
| 714 aggregated_embedding = aggregate_embeddings( | |
| 715 embeddings, | |
| 716 config.pooling_method, | |
| 717 config.use_gpu | |
| 718 ) | |
| 719 | |
| 720 # Set bag metadata. | |
| 721 bsize = bag_data.shape[0] | |
| 722 samples = np.unique(bag_data[:, 0]) | |
| 723 merged_sample_name = ",".join(map(str, samples)) | |
| 724 | |
| 725 # Create row for the bag. | |
| 726 row = { | |
| 727 "sample_name": merged_sample_name, | |
| 728 "bag_label": bag_label, | |
| 729 "split": split, | |
| 730 "bag_size": bsize | |
| 731 } | |
| 732 for j, val in enumerate(aggregated_embedding): | |
| 733 row[f"vector_{j}"] = val | |
| 734 | |
| 735 batch_rows.append(row) | |
| 736 bag_count += 1 | |
| 737 | |
| 738 if len(batch_rows) >= batch_size: | |
| 739 df_batch = pd.DataFrame(batch_rows) | |
| 740 df_batch.to_parquet( | |
| 741 bag_file, | |
| 742 engine="fastparquet", | |
| 743 append=True, | |
| 744 index=False | |
| 745 ) | |
| 746 logging.debug( | |
| 747 "Fixed mode: Wrote a batch of %d rows to %s", | |
| 748 len(batch_rows), | |
| 749 bag_file | |
| 750 ) | |
| 751 batch_rows = [] | |
| 752 del df_batch | |
| 753 gc.collect() | |
| 754 | |
| 755 # Write any remaining rows. | |
| 756 if batch_rows: | |
| 757 df_batch = pd.DataFrame(batch_rows) | |
| 758 df_batch.to_parquet( | |
| 759 bag_file, | |
| 760 engine="fastparquet", | |
| 761 append=True, | |
| 762 index=False | |
| 763 ) | |
| 764 logging.debug( | |
| 765 "Wrote the final batch of %d rows to %s", | |
| 766 len(batch_rows), | |
| 767 bag_file | |
| 768 ) | |
| 769 del df_batch | |
| 770 gc.collect() | |
| 771 | |
| 772 logging.info("Created %d bags for split: %s", bag_count, split) | |
| 773 | |
| 774 else: | |
| 775 # Alternating mode: alternate between labels 0 and 1. | |
| 776 indices_0 = np.where(df_np[:, 1] == 0)[0] | |
| 777 indices_1 = np.where(df_np[:, 1] == 1)[0] | |
| 778 np.random.shuffle(indices_0) | |
| 779 np.random.shuffle(indices_1) | |
| 780 turn = 0 # 0: label 0, 1: label 1. | |
| 781 | |
| 782 while len(indices_0) > 0 or len(indices_1) > 0: | |
| 783 current_bag_size = np.random.randint(bag_min, bag_max + 1) \ | |
| 784 if bag_min != bag_max else bag_min | |
| 785 | |
| 786 if turn == 0: | |
| 787 if len(indices_0) > 0: | |
| 788 num_to_select = min(current_bag_size, len(indices_0)) | |
| 789 selected = indices_0[:num_to_select] | |
| 790 indices_0 = indices_0[num_to_select:] | |
| 791 else: | |
| 792 if len(indices_1) == 0: | |
| 793 break | |
| 794 num_to_select = min(current_bag_size, len(indices_1)) | |
| 795 selected = indices_1[:num_to_select] | |
| 796 indices_1 = indices_1[num_to_select:] | |
| 797 else: | |
| 798 if len(indices_1) > 0: | |
| 799 num_to_select = min(current_bag_size, len(indices_1)) | |
| 800 selected = indices_1[:num_to_select] | |
| 801 indices_1 = indices_1[num_to_select:] | |
| 802 else: | |
| 803 if len(indices_0) == 0: | |
| 804 break | |
| 805 num_to_select = min(current_bag_size, len(indices_0)) | |
| 806 selected = indices_0[:num_to_select] | |
| 807 indices_0 = indices_0[num_to_select:] | |
| 808 | |
| 809 bag_data = df_np[selected] | |
| 810 if bag_data.shape[0] == 0: | |
| 811 break | |
| 812 | |
| 813 # Aggregate embeddings. | |
| 814 vec_col_indices = [ | |
| 815 df.columns.get_loc(col) for col in vector_columns | |
| 816 ] | |
| 817 embeddings = bag_data[:, vec_col_indices].astype(np.float32) | |
| 818 aggregated_embedding = aggregate_embeddings( | |
| 819 embeddings, | |
| 820 config.pooling_method, | |
| 821 config.use_gpu | |
| 822 ) | |
| 823 | |
| 824 # Set bag label and metadata. | |
| 825 bag_label = int(np.any(bag_data[:, 1] == 1)) | |
| 826 bsize = bag_data.shape[0] | |
| 827 samples = np.unique(bag_data[:, 0]) | |
| 828 merged_sample_name = ",".join(map(str, samples)) | |
| 829 | |
| 830 # Create row for the bag. | |
| 831 row = { | |
| 832 "sample_name": merged_sample_name, | |
| 833 "bag_label": bag_label, | |
| 834 "split": split, | |
| 835 "bag_size": bsize | |
| 836 } | |
| 837 for j, val in enumerate(aggregated_embedding): | |
| 838 row[f"vector_{j}"] = val | |
| 839 | |
| 840 batch_rows.append(row) | |
| 841 bag_count += 1 | |
| 842 turn = 1 - turn | |
| 843 | |
| 844 # Write batch to file if batch_size is reached. | |
| 845 if len(batch_rows) >= batch_size: | |
| 846 df_batch = pd.DataFrame(batch_rows) | |
| 847 df_batch.to_parquet( | |
| 848 bag_file, | |
| 849 engine="fastparquet", | |
| 850 append=(bag_count > len(batch_rows)), | |
| 851 index=False | |
| 852 ) | |
| 853 logging.debug( | |
| 854 "Alternating mode: Wrote a batch of %d rows to %s", | |
| 855 len(batch_rows), | |
| 856 bag_file | |
| 857 ) | |
| 858 batch_rows = [] | |
| 859 del df_batch | |
| 860 gc.collect() | |
| 861 | |
| 862 # Write any remaining rows. | |
| 863 if batch_rows: | |
| 864 df_batch = pd.DataFrame(batch_rows) | |
| 865 df_batch.to_parquet( | |
| 866 bag_file, | |
| 867 engine="fastparquet", | |
| 868 append=(bag_count > len(batch_rows)), | |
| 869 index=False | |
| 870 ) | |
| 871 logging.debug( | |
| 872 "Wrote the final batch of %d rows to %s", | |
| 873 len(batch_rows), | |
| 874 bag_file | |
| 875 ) | |
| 876 del df_batch | |
| 877 gc.collect() | |
| 878 | |
| 879 logging.info("Created %d bags for split: %s", bag_count, split) | |
| 880 | |
| 881 | |
| 882 def bag_random(df, split, bag_file, configs, batch_size=500): | |
| 883 """ | |
| 884 Processes the provided DataFrame by randomly selecting instances | |
| 885 to create bags. | |
| 886 """ | |
| 887 logging.info("Processing bag randomly for split %s", split) | |
| 888 | |
| 889 # Identify vector columns (exclude non-vector columns). | |
| 890 vector_columns = [ | |
| 891 col for col in df.columns | |
| 892 if col not in ["sample_name", "label", "split"] | |
| 893 ] | |
| 894 | |
| 895 df_np = df.to_numpy() | |
| 896 | |
| 897 # Create an array of all row indices and shuffle them. | |
| 898 indices = np.arange(df.shape[0]) | |
| 899 np.random.shuffle(indices) | |
| 900 | |
| 901 bag_count = 0 | |
| 902 batch_rows = [] | |
| 903 | |
| 904 # Determine bag size parameters. | |
| 905 if len(configs.bag_size) == 1: | |
| 906 bag_min = bag_max = configs.bag_size[0] | |
| 907 else: | |
| 908 bag_min, bag_max = configs.bag_size | |
| 909 | |
| 910 pos = 0 | |
| 911 total_rows = len(indices) | |
| 912 | |
| 913 # Process until all indices have been used. | |
| 914 while pos < total_rows: | |
| 915 # Ensuring we do not exceed remaining rows. | |
| 916 current_bag_size = (np.random.randint(bag_min, bag_max + 1) | |
| 917 if bag_min != bag_max else bag_min) | |
| 918 current_bag_size = min(current_bag_size, total_rows - pos) | |
| 919 | |
| 920 # Select the indices for this bag. | |
| 921 selected = indices[pos: pos + current_bag_size] | |
| 922 pos += current_bag_size | |
| 923 | |
| 924 # Extract the bag data. | |
| 925 bag_data = df_np[selected] | |
| 926 if bag_data.shape[0] == 0: | |
| 927 break | |
| 928 | |
| 929 # Identify the positions of the vector columns using the column names. | |
| 930 vec_col_indices = [df.columns.get_loc(col) for col in vector_columns] | |
| 931 embeddings = bag_data[:, vec_col_indices].astype(np.float32) | |
| 932 aggregated_embedding = aggregate_embeddings( | |
| 933 embeddings, | |
| 934 configs.pooling_method, | |
| 935 configs.use_gpu | |
| 936 ) | |
| 937 | |
| 938 # Determine bag_label: 1 if any instance in this bag has label == 1. | |
| 939 bag_label = int(np.any(bag_data[:, 1] == 1)) | |
| 940 | |
| 941 # Merge all sample names from the bag (unique names, comma-separated). | |
| 942 samples = np.unique(bag_data[:, 0]) | |
| 943 merged_sample_name = ",".join(map(str, samples)) | |
| 944 | |
| 945 # Use the provided split value. | |
| 946 bag_split = split | |
| 947 bsize = bag_data.shape[0] | |
| 948 | |
| 949 # Build the output row with header fields: | |
| 950 # sample_name, bag_label, split, bag_size, then embeddings. | |
| 951 row = { | |
| 952 "sample_name": merged_sample_name, | |
| 953 "bag_label": bag_label, | |
| 954 "split": bag_split, | |
| 955 "bag_size": bsize | |
| 956 } | |
| 957 for j, val in enumerate(aggregated_embedding): | |
| 958 row[f"vector_{j}"] = val | |
| 959 | |
| 960 batch_rows.append(row) | |
| 961 bag_count += 1 | |
| 962 | |
| 963 # Write out rows in batches. | |
| 964 if len(batch_rows) >= batch_size: | |
| 965 df_batch = pd.DataFrame(batch_rows) | |
| 966 # For the first batch, | |
| 967 # append=False (header written), | |
| 968 # then append=True on subsequent batches. | |
| 969 df_batch.to_parquet( | |
| 970 bag_file, | |
| 971 engine="fastparquet", | |
| 972 append=(bag_count > len(batch_rows)), | |
| 973 index=False | |
| 974 ) | |
| 975 logging.debug( | |
| 976 "Wrote a batch of %d rows to %s", | |
| 977 len(batch_rows), | |
| 978 bag_file | |
| 979 ) | |
| 980 batch_rows = [] | |
| 981 del df_batch | |
| 982 gc.collect() | |
| 983 | |
| 984 # Write any remaining rows. | |
| 985 if batch_rows: | |
| 986 df_batch = pd.DataFrame(batch_rows) | |
| 987 df_batch.to_parquet( | |
| 988 bag_file, | |
| 989 engine="fastparquet", | |
| 990 append=(bag_count > len(batch_rows)), | |
| 991 index=False | |
| 992 ) | |
| 993 logging.debug( | |
| 994 "Wrote the final batch of %d rows to %s", | |
| 995 len(batch_rows), | |
| 996 bag_file | |
| 997 ) | |
| 998 del df_batch | |
| 999 gc.collect() | |
| 1000 | |
| 1001 logging.info("Created %d bags for split: %s", bag_count, split) | |
| 1002 | |
| 1003 | |
| 1004 def imbalance_adjustment(bag_file, split, configs, df): | |
| 1005 """ | |
| 1006 Verifies if the number of bags per label in bag_file is | |
| 1007 within imbalance_cap. | |
| 1008 If not, generates additional bags for the minority label. | |
| 1009 | |
| 1010 Args: | |
| 1011 bag_file (str): Path to the Parquet file containing bags. | |
| 1012 split (str): The current split (e.g., 'train', 'val'). | |
| 1013 config (object): Configuration with imbalance_cap, by_sample, etc. | |
| 1014 df (pd.DataFrame): Original DataFrame for generating additional bags. | |
| 1015 """ | |
| 1016 # Read the bag file and count bags per label | |
| 1017 bags_df = pd.read_parquet(bag_file) | |
| 1018 n0 = (bags_df["bag_label"] == 0).sum() | |
| 1019 n1 = (bags_df["bag_label"] == 1).sum() | |
| 1020 total = n0 + n1 | |
| 1021 | |
| 1022 if total == 0: | |
| 1023 logging.warning("No bags found in %s for split %s", bag_file, split) | |
| 1024 return | |
| 1025 | |
| 1026 # Calculate imbalance as a percentage | |
| 1027 imbalance = abs(n0 - n1) / total * 100 | |
| 1028 logging.info( | |
| 1029 "Split %s: %d bags (label 0: %d, label 1: %d), imbalance %.2f%%", | |
| 1030 split, total, n0, n1, imbalance | |
| 1031 ) | |
| 1032 | |
| 1033 if imbalance > configs.imbalance_cap: | |
| 1034 # Identify minority label | |
| 1035 min_label = 0 if n0 < n1 else 1 | |
| 1036 n_min = n0 if min_label == 0 else n1 | |
| 1037 n_maj = n1 if min_label == 0 else n0 | |
| 1038 | |
| 1039 # Calculate how many bags are needed to balance (aim for equality) | |
| 1040 num_needed = n_maj - n_min | |
| 1041 logging.info( | |
| 1042 "Imbalance %.2f%% exceeds cap %.2f%% in split %s, \ | |
| 1043 need %d bags for label %d", | |
| 1044 imbalance, | |
| 1045 configs.imbalance_cap, | |
| 1046 split, | |
| 1047 num_needed, | |
| 1048 min_label | |
| 1049 ) | |
| 1050 | |
| 1051 # Generate additional bags based on the bag creation method | |
| 1052 if split in configs.by_sample: | |
| 1053 bag_by_sample( | |
| 1054 df, | |
| 1055 split, | |
| 1056 bag_file, | |
| 1057 configs, | |
| 1058 fixed_target_bags=(min_label, num_needed) | |
| 1059 ) | |
| 1060 else: | |
| 1061 bag_in_turns( | |
| 1062 df, | |
| 1063 split, | |
| 1064 bag_file, | |
| 1065 configs, | |
| 1066 fixed_target_bags=(min_label, num_needed) | |
| 1067 ) | |
| 1068 | |
| 1069 # Verify the new balance (optional, for logging) | |
| 1070 updated_bags_df = pd.read_parquet(bag_file) | |
| 1071 new_n0 = (updated_bags_df["bag_label"] == 0).sum() | |
| 1072 new_n1 = (updated_bags_df["bag_label"] == 1).sum() | |
| 1073 new_total = new_n0 + new_n1 | |
| 1074 new_imbalance = abs(new_n0 - new_n1) / new_total * 100 | |
| 1075 logging.info( | |
| 1076 "After adjustment, split %s: %d bags (label 0: %d, label 1: %d), \ | |
| 1077 imbalance %.2f%%", | |
| 1078 split, | |
| 1079 new_total, | |
| 1080 new_n0, | |
| 1081 new_n1, | |
| 1082 new_imbalance | |
| 1083 ) | |
| 1084 else: | |
| 1085 logging.info( | |
| 1086 "Imbalance %.2f%% within cap %.2f%% for split %s, \ | |
| 1087 no adjustment needed", | |
| 1088 imbalance, | |
| 1089 configs.imbalance_cap, | |
| 1090 split | |
| 1091 ) | |
| 1092 | |
| 1093 | |
| 1094 def truncate_bag(bag_file, split): | |
| 1095 """ | |
| 1096 Truncates the bags in the bag_file to balance the counts of label 0 | |
| 1097 and label 1, | |
| 1098 ensuring that the file is never left empty (at least one bag remains). | |
| 1099 | |
| 1100 Args: | |
| 1101 bag_file (str): Path to the Parquet file containing the bags. | |
| 1102 split (str): The current split (e.g., 'train', 'val') | |
| 1103 for logging purposes. | |
| 1104 | |
| 1105 Returns: | |
| 1106 None: Overwrites the bag_file with the truncated bags, | |
| 1107 ensuring at least one bag remains. | |
| 1108 """ | |
| 1109 logging.info("Truncating bags for split %s in file: %s", split, bag_file) | |
| 1110 | |
| 1111 # Step 1: Read the bag file to get the total number of bags | |
| 1112 try: | |
| 1113 bags_df = pd.read_parquet(bag_file) | |
| 1114 except Exception as e: | |
| 1115 logging.error("Failed to read bag file %s: %s", bag_file, e) | |
| 1116 return | |
| 1117 | |
| 1118 total_bags = len(bags_df) | |
| 1119 if total_bags == 0: | |
| 1120 logging.warning("No bags found in %s for split %s", bag_file, split) | |
| 1121 return | |
| 1122 | |
| 1123 # Step 2: Count bags with label 0 and label 1 | |
| 1124 n0 = (bags_df["bag_label"] == 0).sum() | |
| 1125 n1 = (bags_df["bag_label"] == 1).sum() | |
| 1126 logging.info( | |
| 1127 "Split %s: Total bags %d (label 0: %d, label 1: %d)", | |
| 1128 split, | |
| 1129 total_bags, | |
| 1130 n0, | |
| 1131 n1 | |
| 1132 ) | |
| 1133 | |
| 1134 # Determine the minority count and majority label | |
| 1135 min_count = min(n0, n1) | |
| 1136 majority_label = 0 if n0 > n1 else 1 | |
| 1137 | |
| 1138 if n0 == n1: | |
| 1139 logging.info( | |
| 1140 "Bags already balanced for split %s, no truncation needed", | |
| 1141 split | |
| 1142 ) | |
| 1143 return | |
| 1144 | |
| 1145 # Step 3: Adjust min_count to ensure at least one bag remains | |
| 1146 if min_count == 0: | |
| 1147 logging.warning( | |
| 1148 "Minority label has 0 bags in split %s, keeping 1 bag from \ | |
| 1149 majority label %d to avoid empty file", | |
| 1150 split, | |
| 1151 majority_label | |
| 1152 ) | |
| 1153 min_count = 1 # Ensure at least one bag is kept | |
| 1154 | |
| 1155 # Step 4: Truncate excess bags from the majority label | |
| 1156 logging.info( | |
| 1157 "Truncating %d bags from label %d to match %d bags per label", | |
| 1158 max(0, (n0 if majority_label == 0 else n1) - min_count), | |
| 1159 majority_label, | |
| 1160 min_count | |
| 1161 ) | |
| 1162 | |
| 1163 # Shuffle the majority label bags to randomly select which to keep | |
| 1164 majority_bags = bags_df[ | |
| 1165 bags_df["bag_label"] == majority_label | |
| 1166 ].sample(frac=1, random_state=None) | |
| 1167 | |
| 1168 minority_bags = bags_df[bags_df["bag_label"] != majority_label] | |
| 1169 | |
| 1170 # Keep only min_count bags from the majority label | |
| 1171 majority_bags_truncated = majority_bags.iloc[:min_count] | |
| 1172 | |
| 1173 # Combine the truncated majority and minority bags | |
| 1174 truncated_bags_df = pd.concat( | |
| 1175 [majority_bags_truncated, | |
| 1176 minority_bags], | |
| 1177 ignore_index=True | |
| 1178 ) | |
| 1179 | |
| 1180 # Verify that the resulting DataFrame is not empty | |
| 1181 if len(truncated_bags_df) == 0: | |
| 1182 logging.error( | |
| 1183 "Unexpected empty DataFrame after truncation for split %s, \ | |
| 1184 this should not happen", | |
| 1185 split | |
| 1186 ) | |
| 1187 return | |
| 1188 | |
| 1189 # Step 5: Overwrite the bag file with the truncated bags | |
| 1190 try: | |
| 1191 truncated_bags_df.to_parquet( | |
| 1192 bag_file, | |
| 1193 engine="fastparquet", | |
| 1194 index=False | |
| 1195 ) | |
| 1196 logging.info( | |
| 1197 "Overwrote %s with %d balanced bags (label 0: %d, label 1: %d)", | |
| 1198 bag_file, | |
| 1199 len(truncated_bags_df), | |
| 1200 (truncated_bags_df["bag_label"] == 0).sum(), | |
| 1201 (truncated_bags_df["bag_label"] == 1).sum() | |
| 1202 ) | |
| 1203 except Exception as e: | |
| 1204 logging.error("Failed to overwrite bag file %s: %s", bag_file, e) | |
| 1205 | |
| 1206 | |
| 1207 def columns_into_string(bag_file): | |
| 1208 """ | |
| 1209 Reads the bag file (Parquet) from the given path, identifies | |
| 1210 the vector columns (i.e., columns not among 'sample_name', 'bag_label', 'split', | |
| 1211 and 'bag_size'), concatenates these vector values (as strings) into a single | |
| 1212 whitespace-separated string wrapped in double quotes, stored in a new column | |
| 1213 "embeddings", drops the individual vector columns, and writes the modified | |
| 1214 DataFrame back to the same Parquet file. | |
| 1215 | |
| 1216 The final output format is: | |
| 1217 "sample_name", "bag_label", "split", "bag_size", "embeddings" | |
| 1218 where "embeddings" is a string like: "0.1 0.2 0.3" | |
| 1219 """ | |
| 1220 logging.info( | |
| 1221 "Converting vector columns into string for bag file: %s", | |
| 1222 bag_file | |
| 1223 ) | |
| 1224 | |
| 1225 try: | |
| 1226 df = pd.read_parquet(bag_file, engine="fastparquet") | |
| 1227 except Exception as e: | |
| 1228 logging.error("Error reading bag file %s: %s", bag_file, e) | |
| 1229 return | |
| 1230 | |
| 1231 # Define non-vector columns. | |
| 1232 non_vector = ["sample_name", "bag_label", "split", "bag_size"] | |
| 1233 | |
| 1234 # Identify vector columns. | |
| 1235 vector_columns = [col for col in df.columns if col not in non_vector] | |
| 1236 logging.info("Identified vector columns: %s", vector_columns) | |
| 1237 | |
| 1238 # Create new 'embeddings' column by converting vector columns to str, | |
| 1239 # joining them with whitespace, and wrapping the result in double quotes. | |
| 1240 # Use apply() to ensure the result is a Series with one string per row. | |
| 1241 df["embeddings"] = df[vector_columns].astype(str).apply( | |
| 1242 lambda x: " ".join(x), axis=1 | |
| 1243 ) | |
| 1244 # Drop the original vector columns. | |
| 1245 df.drop(columns=vector_columns, inplace=True) | |
| 1246 | |
| 1247 try: | |
| 1248 # Write the modified DataFrame back to the same bag file. | |
| 1249 df.to_parquet(bag_file, engine="fastparquet", index=False) | |
| 1250 logging.info( | |
| 1251 "Conversion complete. Final columns: %s", | |
| 1252 df.columns.tolist() | |
| 1253 ) | |
| 1254 except Exception as e: | |
| 1255 logging.error("Error writing updated bag file %s: %s", bag_file, e) | |
| 1256 | |
| 1257 | |
| 1258 def processing_bag(configs, bag_file, temp_file, split): | |
| 1259 """ | |
| 1260 Processes a single split and writes bag results | |
| 1261 directly to the bag output Parquet file. | |
| 1262 """ | |
| 1263 logging.info("Processing split %s using file: %s", split, temp_file) | |
| 1264 df = pd.read_parquet(temp_file, engine="fastparquet") | |
| 1265 | |
| 1266 if configs.by_sample is not None and split in configs.by_sample: | |
| 1267 bag_by_sample(df, split, bag_file, configs) | |
| 1268 elif configs.balance_enforced: | |
| 1269 bag_in_turns(df, split, bag_file, configs) | |
| 1270 else: | |
| 1271 bag_random(df, split, bag_file, configs) | |
| 1272 | |
| 1273 # Free df if imbalance_adjustment is not needed | |
| 1274 if configs.imbalance_cap is None: | |
| 1275 del df | |
| 1276 gc.collect() | |
| 1277 | |
| 1278 if configs.imbalance_cap is not None: | |
| 1279 imbalance_adjustment(bag_file, split, configs, df) | |
| 1280 del df | |
| 1281 gc.collect() | |
| 1282 elif configs.truncate_bags: | |
| 1283 truncate_bag(bag_file, split) | |
| 1284 | |
| 1285 if configs.ludwig_format: | |
| 1286 columns_into_string(bag_file) | |
| 1287 | |
| 1288 return bag_file | |
| 1289 | |
| 1290 | |
| 1291 def write_final_csv(output_csv, bag_file_paths): | |
| 1292 """ | |
| 1293 Merges all Parquet files into a single CSV file, | |
| 1294 processing one file at a time to minimize memory usage. | |
| 1295 | |
| 1296 Args: | |
| 1297 output_csv (str): Path to the output CSV file specified | |
| 1298 in config.output_csv. | |
| 1299 bag_file_paths (list): List of paths to the Parquet files | |
| 1300 for each split. | |
| 1301 | |
| 1302 Returns: | |
| 1303 str: Path to the output CSV file. | |
| 1304 """ | |
| 1305 logging.info("Merging Parquet files into final CSV: %s", output_csv) | |
| 1306 | |
| 1307 first_file = True # Flag to determine if we need to write the header | |
| 1308 total_rows_written = 0 | |
| 1309 | |
| 1310 # Process each Parquet file sequentially | |
| 1311 for bag_file in bag_file_paths: | |
| 1312 try: | |
| 1313 # Skip empty or invalid files | |
| 1314 if os.path.getsize(bag_file) == 0: | |
| 1315 logging.warning( | |
| 1316 "Parquet file %s is empty (zero size), skipping", | |
| 1317 bag_file | |
| 1318 ) | |
| 1319 continue | |
| 1320 | |
| 1321 # Load the Parquet file into a DataFrame | |
| 1322 df = pd.read_parquet(bag_file, engine="fastparquet") | |
| 1323 if df.empty: | |
| 1324 logging.warning("Parquet file %s is empty, skipping", bag_file) | |
| 1325 continue | |
| 1326 | |
| 1327 logging.info("Loaded %d rows from Parquet file: %s, columns: %s", | |
| 1328 len(df), bag_file, list(df.columns)) | |
| 1329 | |
| 1330 # Write the DataFrame to the CSV file | |
| 1331 # - For the first file, write with header (mode='w') | |
| 1332 # - For subsequent files, append without header (mode='a') | |
| 1333 mode = 'w' if first_file else 'a' | |
| 1334 header = first_file # Write header only for the first file | |
| 1335 df.to_csv(output_csv, mode=mode, header=header, index=False) | |
| 1336 total_rows_written += len(df) | |
| 1337 | |
| 1338 logging.info( | |
| 1339 "Wrote %d rows from %s to CSV, total rows written: %d", | |
| 1340 len(df), bag_file, total_rows_written | |
| 1341 ) | |
| 1342 | |
| 1343 # Clear memory | |
| 1344 del df | |
| 1345 gc.collect() | |
| 1346 | |
| 1347 first_file = False | |
| 1348 | |
| 1349 except Exception as e: | |
| 1350 logging.error("Failed to process Parquet file %s: %s", bag_file, e) | |
| 1351 continue | |
| 1352 | |
| 1353 # Check if any rows were written | |
| 1354 if total_rows_written == 0: | |
| 1355 logging.error( | |
| 1356 "No valid data loaded from Parquet files, cannot create CSV" | |
| 1357 ) | |
| 1358 raise ValueError("No data available to write to CSV") | |
| 1359 | |
| 1360 logging.info( | |
| 1361 "Successfully wrote %d rows to final CSV: %s", | |
| 1362 total_rows_written, | |
| 1363 output_csv | |
| 1364 ) | |
| 1365 return output_csv | |
| 1366 | |
| 1367 | |
| 1368 def process_splits(configs, embedding_files, bag_files): | |
| 1369 """Processes splits in parallel and returns all bags.""" | |
| 1370 splits = [0, 1, 2] # Consistent with setup_temp_files() | |
| 1371 | |
| 1372 # Filter non-empty split files | |
| 1373 valid_info = [] | |
| 1374 for split in splits: | |
| 1375 temp_file = embedding_files[split] | |
| 1376 bag_file = bag_files[split] | |
| 1377 if os.path.getsize(temp_file) > 0: # Check if file has content | |
| 1378 valid_info.append((configs, bag_file, temp_file, split)) | |
| 1379 else: | |
| 1380 logging.info("Skipping empty split file: %s", temp_file) | |
| 1381 | |
| 1382 if not valid_info: | |
| 1383 logging.warning("No non-empty split files to process") | |
| 1384 return [] | |
| 1385 | |
| 1386 # Process splits in parallel and collect bag file paths | |
| 1387 bag_file_paths = [] | |
| 1388 with mp.Pool(processes=mp.cpu_count()) as pool: | |
| 1389 logging.info("Starting multiprocessing") | |
| 1390 bag_file_paths = pool.starmap(processing_bag, valid_info) | |
| 1391 logging.info("Multiprocessing is done") | |
| 1392 | |
| 1393 # Write the final CSV by merging the Parquet files | |
| 1394 output_file = write_final_csv(configs.output_csv, bag_file_paths) | |
| 1395 return output_file | |
| 1396 | |
| 1397 | |
| 1398 def cleanup_temp_files(split_files, bag_outputs): | |
| 1399 """Cleans up temporary Parquet files.""" | |
| 1400 for temp_file in split_files.values(): | |
| 1401 try: | |
| 1402 os.remove(temp_file) | |
| 1403 logging.info("Cleaned up temp file: %s", temp_file) | |
| 1404 except Exception as e: | |
| 1405 logging.error("Error removing %s: %s", temp_file, e) | |
| 1406 for bag_output in bag_outputs.values(): | |
| 1407 try: | |
| 1408 os.remove(bag_output) | |
| 1409 logging.info("Cleaned up temp bag file: %s", bag_output) | |
| 1410 except Exception as e: | |
| 1411 logging.error("Error removing %s: %s", bag_output, e) | |
| 1412 | |
| 1413 | |
| 1414 if __name__ == "__main__": | |
| 1415 mp.set_start_method('spawn', force=True) | |
| 1416 logging.basicConfig( | |
| 1417 level=logging.DEBUG, | |
| 1418 format='%(asctime)s - %(levelname)s - %(message)s' | |
| 1419 ) | |
| 1420 | |
| 1421 parser = argparse.ArgumentParser( | |
| 1422 description="Create bags from embeddings and metadata" | |
| 1423 ) | |
| 1424 parser.add_argument( | |
| 1425 "--embeddings_csv", type=str, required=True, | |
| 1426 help="Path to embeddings CSV" | |
| 1427 ) | |
| 1428 parser.add_argument( | |
| 1429 "--metadata_csv", type=str, required=True, | |
| 1430 help="Path to metadata CSV" | |
| 1431 ) | |
| 1432 parser.add_argument( | |
| 1433 "--split_proportions", type=str, default='0.7,0.1,0.2', | |
| 1434 help="Proportions for train, val, test splits" | |
| 1435 ) | |
| 1436 parser.add_argument( | |
| 1437 "--dataleak", action="store_true", | |
| 1438 help="Prevents data leakage" | |
| 1439 ) | |
| 1440 parser.add_argument( | |
| 1441 "--balance_enforced", action="store_true", | |
| 1442 help="Enforce balanced bagging" | |
| 1443 ) | |
| 1444 parser.add_argument( | |
| 1445 "--bag_size", type=str, required=True, | |
| 1446 help="Bag size (e.g., '4' or '3-5')" | |
| 1447 ) | |
| 1448 parser.add_argument( | |
| 1449 "--pooling_method", type=str, required=True, | |
| 1450 help="Pooling method" | |
| 1451 ) | |
| 1452 parser.add_argument( | |
| 1453 "--by_sample", type=str, default=None, | |
| 1454 help="Splits to bag by sample" | |
| 1455 ) | |
| 1456 parser.add_argument( | |
| 1457 "--repeats", type=int, default=1, | |
| 1458 help="Number of bagging repeats" | |
| 1459 ) | |
| 1460 parser.add_argument( | |
| 1461 "--ludwig_format", action="store_true", | |
| 1462 help="Output in Ludwig format" | |
| 1463 ) | |
| 1464 parser.add_argument( | |
| 1465 "--output_csv", type=str, required=True, | |
| 1466 help="Path to output CSV" | |
| 1467 ) | |
| 1468 parser.add_argument( | |
| 1469 "--random_seed", type=int, default=42, | |
| 1470 help="Random seed" | |
| 1471 ) | |
| 1472 parser.add_argument( | |
| 1473 "--imbalance_cap", type=int, default=None, | |
| 1474 help="Max imbalance percentage" | |
| 1475 ) | |
| 1476 parser.add_argument( | |
| 1477 "--truncate_bags", action="store_true", | |
| 1478 help="Truncate bags for balance" | |
| 1479 ) | |
| 1480 parser.add_argument( | |
| 1481 "--use_gpu", action="store_true", | |
| 1482 help="Use GPU for pooling" | |
| 1483 ) | |
| 1484 args = parser.parse_args() | |
| 1485 | |
| 1486 config = BaggingConfig(args) | |
| 1487 logging.info("Starting bagging with args: %s", config) | |
| 1488 | |
| 1489 set_random_seed(config) | |
| 1490 | |
| 1491 metadata_csv = load_metadata(config.metadata_csv) | |
| 1492 if config.prevent_leakage: | |
| 1493 metadata_csv = split_dataset(metadata_csv, config) | |
| 1494 | |
| 1495 split_temp_files, split_bag_outputs = setup_temp_files() | |
| 1496 | |
| 1497 try: | |
| 1498 logging.info("Writing embeddings to split temp Parquet files") | |
| 1499 distribute_embeddings(config, metadata_csv, split_temp_files) | |
| 1500 | |
| 1501 logging.info("Processing embeddings for each split") | |
| 1502 bags = process_splits(config, split_temp_files, split_bag_outputs) | |
| 1503 logging.info("Bags processed. File generated: %s", bags) | |
| 1504 | |
| 1505 finally: | |
| 1506 cleanup_temp_files(split_temp_files, split_bag_outputs) |
